def __init__(self, params):
# private variables with read access
self.__agent = gym.make(
"racer-v0",
render_mode=params["render_mode"],
sensor_array_type=params["sensor_type"],
sensor_array_params=params["sensor_array_params"])
self.__obs = self.__agent.reset()
self.__done = False
self.__reward = None
self.__network = network.neural_network(
self.agent.observation_space.shape[0], 8)
self.__offline_network = network.neural_network(
self.agent.observation_space.shape[0], 8, requires_grad=False)
self.__memory = None
# public variables
self.render_mode = params["render_mode"]
return
plt.figure(2)
plt.scatter(dataset[:, 0].reshape((dataset.shape[0], 1)),
dataset[:, 3].reshape((dataset.shape[0], 1)))
plt.savefig('./Plots/g.png')
plt.figure(5)
plt.scatter(dataset[:, 0].reshape((dataset.shape[0], 1)),
dataset[:, 4].reshape((dataset.shape[0], 1)))
plt.savefig('./Plots/d.png')
print(dataset.shape)
g = tf.Graph()
with g.as_default():
model_G = neural_network(1, 1, [10], name='Model_G_')
init_g = tf.initialize_all_variables()
x_g = tf.placeholder(tf.float64, [None, 1])
G = model_G.value(x_g)
d1g = model_G.dx(x_g)
d2g = model_G.d2x(x_g)
d = tf.Graph()
with d.as_default():
model_D = neural_network(1, 1, [10, 10], name='Model_D_')
init_d = tf.initialize_all_variables()
x_d = tf.placeholder(tf.float64, [None, 1])
D = model_D.value(x_d)
d1d = model_D.dx(x_d)
d2d = model_D.d2x(x_d)
e_greedy = 0.4
elif 750 <= episode < 800:
e_greedy = 0.3
elif 800 <= episode < 850:
e_greedy = 0.2
else:
e_greedy = 0.1
# sample net structure, train and get accuracy
S, U = table.sample_new_network(epsilon=e_greedy)
while S in Memory_S:
S, U = table.sample_new_network(epsilon=e_greedy)
net_structure = deepcopy(U)
net_structure.append(('T', 4))
nn = neural_network(net_structure=net_structure, input_shape=input_shape)
nn.compile_model()
nn.fit_model(X_train, y_train, batch_size=128, nb_epoch=5, val_X=X_test, val_y=y_test, verbose=1)
accuracy = nn.evaluate_model()
print('score :', accuracy)
# store to replay memory
Memory_S.append(S)
Memory_U.append(U)
Memory_accuracy.append(accuracy)
# update q-table for k times
for memory in range(k_replay_update):
S_sample, U_sample, accuracy_sample = uniform(Memory_S, Memory_U, Memory_accuracy)
table.update_q_values(S_sample, U_sample, accuracy_sample)
lr = 0.01
N = 3
data_sampler = sample_dataset(dataset, batch_size, N)
n_batches = int(len(dataset)/batch_size)
plt.figure(1); plt.scatter(dataset[:,0].reshape((dataset.shape[0],1)), dataset[:,N].reshape((dataset.shape[0],1)))
plt.savefig('./Plots/1.png')
#placeholders for training data
x = tf.placeholder(tf.float64, [None, 1])
y = tf.placeholder(tf.float64, [None, 1])
model = neural_network(1,1,[10], name='Model_G_')
network_out = model.value(x)
loss = tf.reduce_mean(tf.nn.l2_loss(network_out-y))
optimizer = tf.train.AdamOptimizer(learning_rate=lr).minimize(loss)
init = tf.initialize_all_variables()
saver = tf.train.Saver(save_relative_paths=True)
with tf.Session() as sess:
# create initialized variables
best_loss = sys.maxsize
sess.run(init)